Near Duplicate Image Retrieval Based On Geometric Information

With the rapid development of computer and Internet technology, as well as the extensive application of image acquisition equipment, the number of digital images showing explosive growth. In the large number of images, there are a large number of nare duplicate images, the retrieval of similar images has a wide range of application prospects, Such as: the copyright of Internet works of art invasion detection, checking whether same or similar with the exixting trademark, using commodity picture to retrieval similar same goods or to exclude the similar or the same results on retrieval to maintain the diversity of search results, and so on. In recent years, near duplicate image retrieval has gradually become an important branch of image retrieval, which has been paid more and more attention.This thesis also introduce the key technologies need to use on near duplicate image retrieval, and compared scenarios for the use of these technologies, performance and principle. The classical method of image retrieval is to use local features and bag of word model. But the bag of word model only use the describe information of the local characteristics, and no use of geometric information of local features, such as: scale,angle, coordinate. The main content of this paper is to use the geometric information of local feature points to improve the precision of image retrieval system. In this paper, the geometric information of local features is used to improve the performance of image retrieval, and three algorithms are proposed. First proposed the method that consensus guided multiple match removal for geometry verification in image retrieval, considering the correct matching often have more matching consistent with it on transformation parameters space, the method will be uniform partition of the parameter space, ervry time choose the retention of a match from the dense region, and then remove the matching conflicting with it, through constant iteration, leaving only one to one matching. This method can remove a lot of error matching and can be used as a preprocessing step for other geometric calibration methods to improve their performance. The traditional geometric approach is based on the histogram of the parameter space, which can be regarded as a relatively coarse density estimation method.In this paper, a geometric calibration method based on Hough space kernel density is proposed, which is used to calculate the similarity of the density. The experiment resultsshow that this method can improve the performance of image retrieval. Finally, the paper presents a kind of local feature coding method based on geometric information,coding relative of the feature points in Hough space for image retrieval. Experiments show that the method can improve the performance of image retrieval.